Ensure weight transpose is valid for Hopper FP8 training (#1596)

* Update usage of weightmat before saving for backward

Signed-off-by: Guyue Huang <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix for layernorm mlp

Signed-off-by: Guyue Huang <[email protected]>

---------

Signed-off-by: Guyue Huang <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <[email protected]>

Author: 3 people

transformer_engine/pytorch/module/layernorm_linear.py

@@ -341,7 +341,7 @@ def forward(
                 weight.requires_grad and parallel_mode == "column" and sequence_parallel
             )
 
-            # Input with column-wise usage is needed for dgrad GEMM.
+            # Input with column-wise usage is needed for wgrad GEMM.
             if backward_needs_input:
                 if isinstance(ln_out, QuantizedTensor):
                     # For sequence parallel in vanilla FP8, rowwise data is
@@ -350,6 +350,11 @@ def forward(
                     if isinstance(ln_out, MXFP8TensorBase) or not ctx.ln_out_needs_gather:
                         ln_out.update_usage(rowwise_usage=False)
 
+            # Weight with column-wise usage is needed for dgrad GEMM.
+            if inp.requires_grad:
+                if isinstance(weightmat, QuantizedTensor):
+                    weightmat.update_usage(columnwise_usage=True)
+
             if cpu_offloading:
                 if fp8 and weightmat is not None:
                     set_offloading_param(weightmat, "weight_offloading", True)

transformer_engine/pytorch/module/layernorm_mlp.py

@@ -442,6 +442,14 @@ def forward(
             ub_type=tex.CommOverlapType.RS if ub_overlap_rs else None,
             extra_output=rs_out,
         )
+
+        # Weight with column-wise usage is needed for dgrad GEMM.
+        if is_grad_enabled and inp.requires_grad:
+            if isinstance(fc1_weight_final, QuantizedTensor):
+                fc1_weight_final.update_usage(columnwise_usage=True)
+            if isinstance(fc2_weight_final, QuantizedTensor):
+                fc2_weight_final.update_usage(columnwise_usage=True)
+
         if not is_grad_enabled:
             clear_tensor_data(act_out, fc1_out_without_bias, fc1_out)
         else:

transformer_engine/pytorch/module/linear.py

@@ -272,6 +272,11 @@ def forward(
                         inputmat.update_usage(rowwise_usage=False)
                 saved_inputmat = inputmat
 
+            # Weight with column-wise usage is needed for dgrad GEMM.
+            if inp.requires_grad:
+                if isinstance(weightmat, QuantizedTensor):
+                    weightmat.update_usage(columnwise_usage=True)
+
             if cpu_offloading:
                 set_offloading_param(weight, "weight_offloading", True)
                 set_offloading_param(weightmat, "weight_offloading", True)